Most AI persona tools generate a person from a prompt. Candor doesn’t. Every audience and every synthetic participant traces back to specific evidence, and every report comes with the interviews it was drawn from. Here’s what that means in practice and how to read the signals.
A confident-sounding synthetic participant is dangerous if you can’t tell whether their answer is grounded in real-world evidence or generated from thin air. Confidence without provenance is indistinguishable from confabulation. Candor’s answer is to tag what it tells you about an audience and its participants with where it came from, and to never claim more confidence than the evidence supports.
This shows up in two places: every signal in an audience, and every memory field and attribute on a participant, carries a provenance tag. Reports work differently, as the list further down this article explains.
There is something you can check behind this value. Three things land here, listed below. On a mouse, hovering the badge also names which one it was for that field.
Your evidence. The field cites signals pulled out of the research for this audience, your uploaded documents and the sources researched for the study, and at least one of them, checked against the field’s own wording, matches it. The hover says how many of its citations do. A participant whose citations were never checked says Not checked on its sourcing chip, and a cited field there takes its word from the model’s own tag. On a signal, grounded means it is traceable to a specific source that states it, either a fact or quote from your own documents or a page found by web search.
What you told us. The values you fixed when you set the study up. If you gave an age range, an income range, a region, a role, or a company size, the participants carry yours rather than one Candor picked. It also covers what your own words decide: writing “GenZ coffee drinkers” sets an age band, and “nurses in rural settings” sets urban or rural. Leave a box blank and Candor chooses, which reads as Derived instead. Household income is the exception: with no income range, Candor usually picks the figure itself to fit the job, and that reads Assumed. The participant requirements you confirm at setup show on each participant built since this card was added, under What you asked for, and one the participant’s profile meets reads Grounded there. The profile fields that describe the same behaviour still take their word from their own evidence.
The last step of setup shows you those readings before generation starts, so a band you never meant is something you can fix rather than something you discover on a finished profile. Correct one and the field carries your value, with your correction named on the badge. Drop one and Candor stops treating that field as yours, so it reads Derived. The income box shows its reading as you type, as a yearly figure in the study’s currency, and the last step marks a reading that is not a believable yearly household income for the people you described in this region, without changing what you wrote. See setting up your first project.
Public population data. Education and household on U.S. consumer audiences, drawn from the Census American Community Survey for the participant’s region and age band. On a U.S. study built around an occupation, the mix of women and men across the set starts from the share of women published for that occupation in the national employment survey, rather than from a guess at what looks realistic. What you told us about who takes part still comes first, and the mix can move away from the published figure where the study describes a narrower group, such as one seniority level.
One thing Grounded does not mean is a guarantee that a researcher’s belief about their audience is correct. What you told us about who these people are is what they are built to. An age range written as two numbers, such as 25-40, is held in every participant. On a business study, for a participant who comes back outside the company size or below the seniority you set, the fields that miss are corrected, up to twice, and the participant is left out only if it still misses after that, the same as one outside an age range or one who contradicts a requirement you confirmed at setup. On a study with confirmed requirements, the same goes for one who contradicts another box or reading you set, such as region or industry. What you believe about how they think and decide is never treated as evidence, because that is what the research exists to test. A requirement you confirm is different: it decides who takes part, which is why it reads Grounded under What you asked for, but it is still not evidence for the profile fields.
Built from your evidence, the audience you described, or population data, but no single source states this field’s value. Two things produce it. Fitting a value to the kind of participant your research described, where you did not set it yourself: where you left a setup box blank, Candor picked a value that fits the audience you described (except a household income Candor picked with no income range, which reads Assumed). And reasoning a step past what the research says outright: the evidence points here and may even be cited, but nothing in it states this value. Where Candor checks a field’s citations, each is tested against the field’s own wording, so related evidence that does not state the value reads Derived, whatever word the model chose.
Weaker than grounded, because the connection is Candor’s rather than a source’s. Stronger than a guess, because there is evidence underneath it. On an audience, a derived signal is not lifted from a single stated source. Most are several sources summarized into one signal. Some are a document author’s own reading of their evidence: a survey memo’s “likely because they trust the existing process”, or an interview summary written in the analyst’s voice rather than the interviewee’s. Candor keeps those as hypotheses: they are left out of the evidence a participant is written from, and cannot on their own define a group or a dimension. Others land there because Candor’s own accuracy check found the claim says more than its sources support, or because no source was recorded at all. So a derived signal can still show source chips beside it.
That accuracy check is worth knowing about. A signal built from several sources is re-read against their wording, because pulling findings together is where a claim is most likely to bend: a number that measured one thing described as another, a study’s country or setting dropped, or two figures with different bases merged into one range. A claim that doesn’t hold up gets one automatic attempt at a rewrite during the run, kept only if re-checking rates it better. Anything still flagged afterwards reads Derived rather than Grounded. Nothing is removed either way. On an ordinary run the page shows nothing about the check; a line above the signal list appears only when it did not run, or found nothing it could trace, to say those signals are unverified.
“No single source” appears beside some derived signals, and it is a statement about the trail rather than about the claim. It means no source was recorded for that signal, which is normal for a claim drawn from general research rather than from anything found for your audience. Occasionally it shows on a signal that WAS built from named sources, because Candor could not match its citations back to the pages it stored. That is a gap in the trail, not a claim without a basis, and you can still check it against the grounded signals on the same theme.
Candor is filling in the gap based on everything it knows, without being able to trace this to any specific data points. Often a reasonable read, but not something your research showed. Marked this way deliberately, not by accident. Treat assumed claims as things to test, not things to trust.
Candor used to show five, splitting Grounded and Derived by the mechanism that produced a value: whether it came from a distribution, from the pattern of a group, or from a citation. That told you how the pipeline worked rather than what you could do about it, and the two extra words were the two nobody could define without the legend.
Three answers the question a researcher actually has: can I check this? The mechanism is still recorded on every value, and nothing was retagged, so a participant generated last year reads the same way as one generated today.
The tags are also colour-coded in the UI so you can scan a participant profile and see at a glance how much of it you can check, how much follows from the evidence, and how much is flagged as filled in.
Three habits that pay off:
Grounding tells you where a claim came from. It doesn’t tell you whether the source was correct. Web evidence can be wrong. Uploaded research can be incomplete. A grounded participant built from biased sources will reflect that bias. Provenance is a tool for healthy scepticism, not a replacement for it.
One narrower check does run on what a participant says. After each interview, Candor reads the transcript for claims about how the world works and has a stronger model flag the ones that look wrong. It then looks each flagged claim up on the web, and records which of that participant’s stored facts and story details state a claim published sources contradict. Nothing is taken away. Candor used to remove those details, and stopped in September 2026, because measuring the step that picks them showed it names true personal stories alongside the wrong ones. The transcript itself is never edited, the check finds some errors and misses others, and nothing in the product shows its result yet.
Look at the provenance tag. Grounded means there is something you can check behind it that fits the field: a cited signal from your audience evidence that matches the field's wording where Candor checked it, a value you fixed yourself when you set the study up, or public population data. Hovering the badge names which one it was. Derived means the value was built from your evidence, the audience you described, or population data, but no single source states it: either a value Candor picked to fit the audience where you left a setup box blank, or a step past what the research says outright, where the evidence may be cited but nothing in it states the value. Assumed means the model is filling the gap from everything it knows, without being able to trace the value to any specific data points, and marks it that way deliberately so you know. A grounded decision rule and an assumed one shouldn't carry the same weight in your head.
Provenance tracks where a claim came from, not whether the source was correct. Web evidence can be wrong. Uploaded research can be incomplete or biased. A grounded synthetic participant built from biased sources will reflect that bias. That's why provenance is a tool for healthy scepticism, not a replacement for it. The fix is to click through on grounded signals you find surprising and verify the source. Strong evidence sources produce strong participants. Weak sources produce weak ones, no matter how confident the model sounds.
Candor used to show five, splitting the top two by the mechanism that produced a value: whether it came from a distribution, from the pattern of a group, or from a citation. That described how the pipeline works rather than what you can do about it, and the two extra words were the two nobody could define without the legend. Three answers the question a researcher actually has: can I check this? The mechanism is still recorded on every value, and nothing was retagged, so a participant generated last year reads the same way as one generated today.
Yes, for the facts and quotes your documents state. Those get higher priority than web sources when the audience is built. The reasoning: you uploaded those documents because they're directly relevant to your audience, while web search casts a wide net and pulls in adjacent or partial matches. One distinction matters: when a document's author interprets their own evidence (a survey memo saying a group is hesitant 'likely because they trust the existing process', or an interview summary written in the analyst's voice), Candor records that sentence as the author's reading, not as a fact. It shows as Derived on the signal list, it is left out of the evidence participants are written from, and it cannot on its own define a group or a dimension. The audience-quality score on each synthetic participant reflects the stated evidence. The more of your audience is built from facts in documents you trust, the higher the quality score, and the more confident you can be in that participant's claims.
Synthesized signals, the ones built from several sources at once rather than lifted from a single one, get re-read against the wording of the sources they cite. Pulling two findings together is where a claim is most likely to bend: a number that measured one thing described as another, a study’s country or care setting dropped, or two figures with different bases merged into one range. A claim that doesn’t hold up gets one automatic attempt at a rewrite during the run, against its own sources, kept only if re-checking rates it better. Nothing is removed either way. On an ordinary run the page shows nothing about the check. A line above the signal list appears only when the check did not run, or found no synthesized signal it could trace back to its sources, and it says those signals are unverified. Read it when it appears, because an unchecked list looks identical to one where every claim held up.
Two places. On every audience signal (a coloured dot plus a source pill for each source behind it, linking out for web pages and naming the file for uploads). On every synthetic participant's memory field and B2C/B2B attribute (a provenance badge next to the value). Synthesis reports carry no provenance tags. Today's reports show their evidence through participant counts, named paraphrases in validation and concept-test reports, and the full interview transcripts that come with every report. Scanning provenance is the fastest way to read a participant profile critically. The colour-coded tags let you see at a glance how much of it you can check, how much follows from the evidence, and how much is assumed.
Start with a free project, or walk through Candor with us first.